You are here

Automatic Blood Smear Analysis with Artificial Intelligence and Smartphones.

Printer-friendly versionPrinter-friendly version
Yu H, Yang F, Silamut R, Maude S, Jaeger S, Antani SK
ASTMH 68th Annual Meeting, Washington DC, Nov. 20-24, 2019.
Abstract: 

MalariaScreener is the first software using artificial intelligence, in particular deep learning, to process both thin and thick blood smear images on smartphones. When running on a phone attached to the eyepiece of a microscope, MalariaScreener can detect parasites in thick smear images captured by the smartphone's camera. It can also discriminate between infected and uninfected red blood cells in thin blood smear images. The idea is to provide a cost-effective alternative for malaria diagnosis in resource-limited regions that does not depend on expert knowledge. This could also help in standardizing parasite counts, which often depend on the experience and skill of the microscopist. We trained our system on hundreds of thousands of manually annotated parasites and blood cells, in images from hospitals and clinics in Bangladesh and Thailand, so that it captures the typical shape and appearance of parasites and cells. The trained system can detect parasites in thick smears and discriminate between infected and uninfected red blood cells in thin smears. For thick smears, we measure an area under the ROC curve (AUC) of 85% on patient level. For thin smears, using ensemble techniques, we achieve an AUC of 99.92% for cell classification (infected/uninfected). The underlying algorithms utilize customized convolutional neural network models (CNNs). An embedded database system can store patient data, screening results, disease severity, and patient histories. Moreover, researchers can save blood smear images captured during the screening process together with their manual counts of parasites and blood cells, which can then be used later for training to improve future automated screening. The internal algorithms have been optimized for fast processing speed and leave a small memory footprint. The user interface allows easy access to the image data and review of all stored information. MalariaScreener is currently being field-tested in several countries. Due to its modular and flexible design, Malaria Screener could be extended to support screening of other blood-based diseases and beyond, when given the properly trained machine learning models.

Yu H, Yang F, Silamut R, Maude S, Jaeger S, Antani SK. Automatic Blood Smear Analysis with Artificial Intelligence and Smartphones. ASTMH 68th Annual Meeting, Washington DC, Nov. 20-24, 2019.